Tabular Foundation Models Are Competitive Cellular Perturbation Predictors Across Biological Scales
Predicting how cells respond to genetic and chemical perturbations is a central challenge in drug discovery and functional genomics. A growing ecosystem of specialized single-cell foundation models has been developed to address this problem, yet their practical advantage over domain-agnostic approaches remains unclear. Here we evaluate the power of Tabular Foundation Models such as TabICL and TabPFN, general-purpose pre-trained regression models, against domain-specific architectures including PRESAGE, scGPT, scLAMBDA, STACK and Prophet across four complementary evaluation settings: cell-level in-context cross-cell-type prediction, pseudobulk perturbation prediction on five Perturb-seq datasets of cell-lines, a genome-wide CRISPR screen in primary human CD4+ T cells, and embryo-level cell-type composition prediction in a zebrafish developmental perturbation atlas. In the cell-level cross-cell type perturbation prediction, Tabular Foundation Models perform on par or better than specialized models. On pseudobulk perturbation prediction, Tabular Foundation Models consistently out-perform specialized baselines across multiple evaluation metrics and datasets. On whole-embryo cell-type composition prediction, Tabular Foundation Models are competitive with specialized baselines. These results demonstrate that general-purpose tabular in-context learning provides a strong and scalable alternative to bespoke biological architectures for perturbation response modeling across cell systems and scales.